Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Class Semantic Attribute Perception Guided Zero-Shot Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CSAP-ZSL, an end-to-end zero-shot image classification method combining contrastive coarse-grained attribute alignment with soft region partitioning and region-partition-based fine-grained attribute alignment on a ViT backbone.
Problem
Zero-shot image classification requires aligning image features with class semantic attributes, but existing overall- and part-embedding methods fail to localize individual attributes within image regions and ignore relations between regions.
Approach
Splits an image into regions, softly clusters similar spatially-neighboring regions via a graph-cut-style region partitioning network, and aligns each cluster's max response with attribute prototypes, combined with a coarse contrastive loss between the overall image embedding and class attribute embeddings, trained end-to-end.
Key results
- CSAP-ZSL reaches H = 74.6 on AWA2, 70.6 on CUB, and 52.2 on SUN in generalized ZSL, the best among compared methods in Table 2.
- On SUN, H improves from 45.8 (DUET) to 52.2, a gain of about 6.4 points.
- Ablation: fine-tuning the encoder end-to-end (M2) raises AWA2 H from 64.7 to 68.2 over frozen features (M1), and the full model reaches H 74.6 on AWA2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune the feature extraction network (ablation) | CSAP-ZSL coarse-grained module · AWA2 | H: 64.7 -> 68.2 | Helped |
| fine-tune the feature extraction network (ablation) | CSAP-ZSL coarse-grained module · CUB | H: 59.8 -> 65.1 | Helped |
| fine-tune the feature extraction network (ablation) | CSAP-ZSL coarse-grained module · SUN | H: 45.9 -> 48.9 | Helped |
| add the fine-grained perception module without soft region partitioning (ablation) | CSAP-ZSL · AWA2 | H: 68.2 -> 71.1 | Helped |
| add the fine-grained perception module without soft region partitioning (ablation) | CSAP-ZSL · CUB | H: 65.1 -> 67.7 | Helped |
| add the fine-grained perception module without soft region partitioning (ablation) | CSAP-ZSL · SUN | H: 48.9 -> 50.0 | Helped |
| add soft region partitioning to the coarse- and fine-grained modules (ablation) | CSAP-ZSL · AWA2 | H: 71.1 -> 74.6 | Helped |
| add soft region partitioning to the coarse- and fine-grained modules (ablation) | CSAP-ZSL · CUB | H: 67.7 -> 70.6 | Helped |
| add soft region partitioning to the coarse- and fine-grained modules (ablation) | CSAP-ZSL · SUN | H: 50.0 -> 52.2 | Helped |
Limitations
- Hyperparameters (alpha, beta, K, radius, gamma) are tuned per dataset, with a different search per dataset.
- Compared against a mix of end-to-end and non-end-to-end methods, and the paper does not report compute cost.
- Evaluated only on three standard ZSL benchmarks (AWA2, CUB, SUN).
Experimental setup
Datasets: AWA2 CUB SUN
Benchmarks: AWA2 CUB SUN
Metrics: top-1 accuracy on seen classes (S) top-1 accuracy on unseen classes (U) harmonic mean (H)
Baselines: DUET TransZero++ MSDN GNDAN TransZero DAZLE SCE-GZSL ICCE mVACA CE-GZSL
Builds on: TransZero APN DAZLE GNDAN
Closest papers by meaning
- DUET: Cross-Modal Semantic Grounding for Contrastive Zero-Shot Learning (AAAI 2023 · similarity 0.91)
- Fine-Grained Generalized Zero-Shot Learning via Dense Attribute-Based Attention (CVPR 2020 · similarity 0.91)
- Visual and Semantic Prompt Collaboration for Generalized Zero-Shot Learning (CVPR 2025 · similarity 0.89)
- Epsilon: Exploring Comprehensive Visual-Semantic Projection for Multi-Label Zero-Shot Learning (AAAI 2025 · similarity 0.89)
- Super-Class Guided Transformer for Zero-Shot Attribute Classification (AAAI 2025 · similarity 0.88)
- Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.